fix(agent): recover interrupted browser tool turns

[DavidBabinec]

What changed

• tie each chat turn to response-stream cancellation so reloads, disconnects, and failed enqueues abort the provider request, browser bridge, and conversation writer lock
• treat a rejected browser bridge as one terminal transport failure: persist/emit one failed tool result, preserve already-billed usage, and stop instead of asking the model to retry against the same dead bridge
• keep resolved { ok: false } browser-tool outcomes recoverable by the model
• propagate active tool-result POST failures, distinguish them from intentional Stop, detect truncated NDJSON streams, and finalize every pending live tool badge
• rehydrate unmatched or malformed persisted screenshot calls as interrupted errors without reconstructing session-only preview images
• reconnect workspace MCP bridges with a fresh per-attempt controller when result delivery fails, while preserving protocol-level MCP tool errors
• share interrupted-call recovery between browser history and provider replay, update the lifecycle documentation, and add focused regressions

Why

A server restart replaces the in-memory browser-bridge registry. The browser could finish capturing a screenshot, then POST its result to a new runtime that no longer owned the old request. The client previously ignored that 404, while the old turn waited 90 seconds and the model repeatedly retried the screenshot. This makes the stale turn terminate once, preserves the failed round's usage/cost, and gives both live and reloaded conversations an honest terminal state.

Verification

• bun test
• bun run build
• bun run lint
• bunx react-doctor --verbose --diff origin/main --no-score
• focused lifecycle regressions across client rehydration/delivery, tool loop, chat cancellation, runner persistence, history replay, and MCP bridge recovery